Artificial intelligence software: FDA news
In the American regulatory framework, artificial intelligence (AI) and machine learning-(ML)based technologies, that are intended to be used for a single, or more, medical purposes (i.e., to treat, detect, cure, reduce or prevent diseases) are included in the category of software as a medical device (SaMD). It should be noticed that, according to the definition given by the International Medical Device Regulators’ Forum, SaMD are those software packages that are intended for use for one or more medical purposes, and that perform these purposes without being part of a medical hardware device. Specifically, AI/ML-based SaMD differ from other SaMDs due to their ability to learn from real-world feedback and, therefore, can improve their performance.
These technologies have an enormous potential that is related to the development of healthcare, i.e., enabling the possibility for early and more detailed diagnosis.
FDA discussion paper
From a regulatory perspective, there is a need to prepare an adequate regulatory framework in order to support the constant learning and updating of the algorithms that are the basis of the software and, at the same time, to ensure appropriate levels of safety and effectiveness over the entire device’s lifecycle. With this purpose, on April 2, 2019, the FDA published a discussion paper on a proposal for a regulatory framework for modifications to AI/ML-based software.
The application of the total product lifecycle regulatory approach to AI/ML based SaMDs
The discussion paper’s legislative approach considers the whole lifecycle of AI/ML based medical devices (Total product lifecycle regulatory approach), with the purpose of ensuring the products’ safety and quality from their premarket development to the necessary after sales services. It is a similar approach to that envisaged by the software pre-certification program, published by the FDA, in January 2019. This regulation seems to be applicable only to those AI/ML-based SaMDs that require a premarket submission.
In order to achieve this purpose, the software manufacturer has to establish quality and good practice standards inside his structure for the entire lifecycle of the product (Quality systems and good machine learning practices – Gmlp), in order to achieve the clinical validation of the software.
Predetermined software change control plan
Secondly, a verification of safety and effectiveness requirements during the premarket phase of the product is requested. For this purpose, the discussion paper has proposed a pre-determined change control plan which can be used during the initial premarket request for an AI/ML based SaMD. The predetermined change control plan should include the types of modifications that are to be expected as a result of the constant updating and retraining of such software (SaMD pre-specifications) and the associated methodology should be used in order to implement those changes in a controlled manner, while managing risks to patients (Algorithm Change Protocol).
Regulation of AI/ML-based SaMD modifications after they are put on the market
Possible modifications, due to the software’s adaptation to the real-world, are considered to be changes to SaMD when it has already received the FDA’s authorization of the premarket submission. The FDA discussion paper requires manufacturers to implement these changes, taking into account the risk for the patients.
In the light of the regulatory framework described above, if the changes that are made fall within the limits of the modifications already approved in the pre-determined change control plan, the software manufacturer should simply document the modification made on the software. Otherwise, manufacturers will have to submit a new market application to the FDA.
Manufacturers’ transparency and monitoring duties
Finally, the manufacturer should commit to the principles of transparency and real-word performance monitoring, providing reports on the real performances of the devices during their post-market phase, ensuring that the labelling describes, in a suitable and detailed manner, those modifications that have been made on the software.
In conclusion, the FDA discussion paper affirms that the new regulatory approach should provide for a reasonable assurance of safety and effectiveness in relation to AI/ML-based SaMDs during the entire software lifecycle. Furthermore, this regulatory approach would enable both the FDA and the manufacturers to test and monitor these devices from the pre- to the post-market phase.